Software Inversion
The hard part moved from “can we build it?” to “do we know what to build?”
The old applied ML stack required specialized expertise: feature engineering, model training, evaluation pipelines, and tuning. That work still matters in some domains, but it is no longer the default starting point for most business AI.
LLMs inverted the first move. For many use cases, the model is available as an API. The harder questions are product and operations questions:
- What work should this improve?
- What context does it need?
- What should it produce?
- Where does human judgment remain essential?
- How should it fail visibly and safely?
These are software questions. They are workflow questions. They are taste questions.
The New Bottleneck
“Your engineers can build AI” is no longer a differentiated claim. The question has shifted from whether AI can be added to whether the organization knows what should be encoded, where it should live, and how it should be governed.
The next inversion is that AI is not only a component in the software. It is increasingly part of the team that builds the software. That makes judgment even more important, not less.
If building gets cheaper, choosing well gets more valuable.
Implication
Shipping AI features is not the goal. Building operating capability is. The constraint has moved from “can we build?” to “do we have coherent judgment worth encoding?”
Contrarian To
“AI projects require specialized ML teams”
Sometimes they do. Most applied business AI starts somewhere else: with workflow knowledge, product judgment, context, and the ability to build useful software around a model.